high performance computing on graphics processing units: hgpu.org

In this work, we propose the implementation of a 3D object recognition system using Convolutional Neural Networks. For that purpose, we first analyzed the theoretical foundations of that kind of neural networks. Next, we discussed ways of representing 3D data in a compact and structured manner to feed the neural network. Those representations consist of a grid-like structure (fixed and adaptive) and a measure for the occupancy of each cell of the grid (binary, normalized point density, and surface intersection). At last, 2.5D and 3D Convolutional Neural Network architectures were implemented and tested using those volumetric representations. The experimentation included an in-depth study of their performance in synthetically simulated adverse conditions that characterize the real-world, i.e., noise and occlusions. The resulting system, the best one out of that experimentation, is able to efficiently recognize objects in three dimensions with a success rate of 85% in a common household CAD objects dataset.